Permutation-Based Causal Structure Learning from Interventional Data with Unknown Targets

Recent innovations in gene editing and gene sequencing technologies have opened the door to developing a more complete understanding of gene regulatory networks, with applications for disease diagnosis, drug development, and biochemical engineering. In particular, the wealth of interventional data from gene knockout experiments presents a ___ source of information for learning causal models of gene regulatory networks directly from gene expression data. However, most existing causal inference methods for learning from interventional data share a common flaw: all intervention targets are assumed to be known in advance. Since gene knockout experiments are known to have off-target effects, these methods may fail on real data. We propose a new method, the Unknown-Target Interventional Greedy Sparsest Permutation (UT-IGSP) algorithm, with consistency guarantees for learning causal DAGs and intervention targets from interventional data with off-target effects. In this talk, I will describe our identifiability results for interventional data with unknown intervention targets, and describe the UT-IGSP algorithm for learning causal models from such data.

Joint work with Yuhao Wang and Caroline Uhler.